Auxiliary-wrapper guided and multi-criteria filter based evolutionary band selection algorithm for hyperspectral image classification

Abstract

The abundant bands contained in hyperspectral images (HSIs) may lead to significant challenges to subsequent data distribution and processing. Band selection (BS), as a dimensionality reduction technique, can efficiently mitigate the data volume of HSIs and retain the physical meaning in bands. Traditional filter based BS techniques often employ various metrics to assess the effectiveness of the selected band subset, rather than utilizing actual classification performance on the classifiers. However, these metrics often do not align with the actual classification performance. To tackle this problem, we propose the auxiliary-wrapper guided and multi-criteria filter based evolutionary BS algorithm for hyperspectral image classification, which integrates an auxiliary wrapper into the filter based BS process to select high-quality band subsets by combining filter metrics from a large number of unlabeled samples and real classification performance from a few labeled samples. Firstly, hyperspectral BS is formulated as a triple-objective optimization problem to evaluate the subsets of bands from multiple perspectives. Moreover, the auxiliary wrapper, where the classification performance of the selected bands is evaluated using only a few labeled samples and a basic classifier, is introduced to further guide the triple-objective optimization in the filter based BS. To keep the size of the selected bands stable in the evolutionary process, the leader-based learning strategy is designed, leveraging the transferring of the bands selected by the wrapper task to the filter task and further inside the filter task in a hierarchical manner. Experimental results on different standard HSI datasets show that the proposed WFBS method can achieve better band subsets compared with the existing unsupervised, semi-supervised and even deep learning based BS methods.